An Automatic Premature Ventricular Contraction Recognition System Based on Imbalanced Dataset and Pre-Trained...

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Signal Processing
• Wearable Health Technology

Background:

• Accurate detection of Premature Ventricular Contractions (PVCs) from electrocardiograms (ECGs) is vital for diagnosing heart failure.
• Manual ECG analysis is time-consuming and requires significant cardiologist effort.
• Advancements in wearable sensors necessitate automated, patient-specific diagnostic systems.

Purpose of the Study:

• To develop and evaluate an automated deep learning method for precise PVC detection using a pre-trained ResNet-18 model.
• To leverage transfer learning to overcome the need for large training datasets and complex feature engineering.
• To assess the method's efficacy and generalizability on established arrhythmia datasets.

Main Methods:

• Utilized a pre-trained ResNet-18 deep residual network with a transfer learning mechanism for automatic feature extraction.
• Segmented ECG beats using the Pan-Tompkins algorithm and converted them into 2D images for model input.
• Optimized the model using weighted random samples, on-the-fly augmentation, Adam optimizer, and callback features.

Main Results:

• Achieved high detection accuracies of 99.93% on the MIT-BIH dataset and 99.77% on the INCART dataset using leave-one-subject-out cross-validation.
• Demonstrated superior performance compared to state-of-the-art methods on unseen data.
• Showcased satisfactory results without complex preprocessing or model design.

Conclusions:

• The proposed deep learning method offers an effective and generalizable approach for automated PVC detection from ECGs.
• The transfer learning strategy effectively handles data limitations and model complexity.
• This automated system shows promise for real-world cardiac monitoring using diverse wearable sensor devices.